Download Aalborg Universitet

Aalborg Universitet
Artificial Neural Network Algorithm for Condition Monitoring of DC-link Capacitors
Based on Capacitance Estimation
Soliman, Hammam Abdelaal Hammam; Wang, Huai; Gadalla, Brwene Salah Abdelkarim;
Blaabjerg, Frede
Published in:
Journal of Renewable Energy and Sustainable Development (RESD)
Publication date:
2015
Document Version
Accepted manuscript, peer reviewed version
Link to publication from Aalborg University
Citation for published version (APA):
Soliman, H. A. H., Wang, H., Gadalla, B. S. A., & Blaabjerg, F. (2015). Artificial Neural Network Algorithm for
Condition Monitoring of DC-link Capacitors Based on Capacitance Estimation. Journal of Renewable Energy and
Sustainable Development (RESD), 1(2), 294-299.
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
? Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
? You may not further distribute the material or use it for any profit-making activity or commercial gain
? You may freely distribute the URL identifying the publication in the public portal ?
Take down policy
If you believe that this document breaches copyright please contact us at [email protected] providing details, and we will remove access to
the work immediately and investigate your claim.
Downloaded from vbn.aau.dk on: September 17, 2016
Journal of Renewable Energy and Sustainable Development (RESD)
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
Artificial Neural Network Algorithm for Condition
Monitoring of DC-link Capacitors Based on
Capacitance Estimation
Hammam Soliman, Huai Wang, IEEE Member, Brwene Gadalla, Frede Blaabjerg, IEEE Fellow
Department of Energy Technology, Aalborg University,
Aalborg 9220, Denmark,
[email protected], [email protected], [email protected], [email protected]
Abstract - In power electronic converters, reliability
of DC-link capacitors is one of the critical issues.
The estimation of their health status as an
application of condition monitoring have been an
attractive subject for industrial field and hence for
the academic research filed as well. More reliable
solutions are required to be adopted by the industry
applications in which usage of extra hardware,
increased cost, and low estimation accuracy are
the main challenges. Therefore, development of
new condition monitoring methods based on software
solutions could be the new era that covers the
aforementioned challenges.
A capacitance
estimation
method
based
on Artificial Neural
Network (ANN) algorithm is therefore proposed in this
paper.
The implemented ANN estimated back
converter. Analysis of the error of the capacitance
estimation
is also given. The presented method
enables a pure software based approach with
high parameter estimation accuracy.
Keywords - Capacitor condition monitoring; capacitor
health status; Capacitance estimation.
In the last two decades, a large number of research
results on condition monitoring of capacitors has been
published. The majority of the condition monitoring
methods for capacitors are based on estimation of the
capacitance C and equivalent series resistance
(ESR), which are indicators of the degradation of
capacitors [3]. For aluminum electrolytic capacitors,
the widely accepted end-of-life criteria are 20%
capacitance reduction or double of the ESR. For film
capacitors, a reduction of 2% to 5% capacitance may
indicate the reach of end-of-life. Therefore, a method
which can estimate the C value could be applied on
both aluminum electrolytic and film capacitors.
However, obtaining the values of C or ESR is an
important step since it gives an indication of the
ageing process and its acceleration. Fig. 2(a) shows a
simplified equivalent model of capacitors and Fig. 2(b)
plots the corresponding frequency characteristics. It
can be noted that the capacitor impedances are
distinguished by three frequency regions dominated
by capacitance, ESR and the Equivalent Series
Inductance (ESL), respectively.
I. INTRODUCTION
Condition monitoring is an important strategy to
estimate the health condition of power electronic
components, convert- ers and systems. It is widely
applied in reliability or safety critical applications, such
as wind turbines, electrical aircraft, electric vehicles,
etc., enabling the indication of future failure
occurrences and preventive maintenances. In [1], the
condition monitoring of semiconductor devices used
in power electronics is well reviewed. Besides the
power devices, and according to [2], electrolytic
capacitors are sharing 60% of the failure distribution
for power converter elements as shown in Fig.1,
therefore, capacitors are another type of reliability
critical components.
Fig .1. Distribution of failure for power converetr elements [2
C
ESR
ESL
(a) Simplified equivalent model of capacitors
294
Journal of Renewable Energy and Sustainable Development (RESD)
(b) Impedance characteristic of capacitors
Fig .2.
Equivalent model and impedance characteristics of
capacitors.
From the methodology point of view in [4],
condition monitoring methods in the literature are
classified into three categories as the following: a)
Capacitor ripple current sensor based methods, b)
Circuit model based methods, and c) Data and
advanced algorithm based methods. The following
three subsections are defining the principle of each
category with its corresponding examples from the
literature.
A. Capacitor ripple current sensor based methods
The basic principle of the this category is to estimate
the capacitance and/or the ESR by using the
capacitor ripple voltage and current information at
region I and II (as shown in Fig. 2(b)). In [2, 5–7], an
external current injection at low fre-quency is the
main approach to achieve condition monitoring, it has
been applied to PWM AC/DC/AC converter in [2], [5],
and [7], while in [6] it has been applied to a submodule capacitor in moulder multilevel converter.
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
estimated is the main concept. In [9], an external
voltage is injected to the reference voltage of the
capacitor at low frequency, the obtained capacitor
power is used as a training data in sake of finding an
identification model based on Support Vector
Regression (SVR). After using a group of training
data, a generated function is used to analyse the
correlation between the known capacitor power and
its corresponding capacitance value. Although the
previous mentioned methods have been ver- ified by
simulation and experimental work, errors, complexity, and cost increasing due to extra hardware are
common shortcomings. Therefore, the developed
technologies are rarely adopted in practical industry
applications, implying that new condition monitoring
methods based on software solutions and existing
feedback signals, without adding any hardware cost,
could be more promising in practical applications.
This paper aims to propose a condition monitoring
method based on Artificial Neural Network (ANN) that
uses existing power stage and control information
and existing spare re- sources of digital controllers. It
requires no extra hardware cir- cuitry (e.g., current
sensors and corresponding signal condition circuits),
no external signal injection, and therefore minimises
the increased complexity and cost. Main sections in
this paper are as the following: Section II gives the
basic principle of ANN applied for capacitor condition
monitoring. Section III
B. Circuit model based methods
This category is based on that instead of injecting
external signals, the capacitor current can be
obtained indirectly de- pending on both the circuit
model and the operation principle of PWM switching
converters. In [8], an on-line condition monitoring
based on capacitance estimation is proposed, the
capacitor ripple current is calculated using the
difference between the input current sensor, and the
output current flows to the inverter which is based on
the transistor switching statues.
Fig .3.
illustrates the applied ANN to a back to back
converter study case. Section IV presents the results
achieved by the proposed method based on ANN,
followed by the conclusion.
II.
C. Data and advanced algorithm based methods
This category, obtaining a strong correlation between
the available parameters and the parameters to be
The structure of the Artificial Neural Network.
ANN FOR CAPACITOR CONDITION
MONITORING
Implementation of capacitor condition monitoring
295
Journal of Renewable Energy and Sustainable Development (RESD)
using ANN is motivated by the shortcomings that
have been investigated earlier in this paper. Avoiding
the usage of direct/indirect current sensors is one of
the main advantages of using ANN to obtain the
estimated value of C. Instead of sensing the capacitor
current iC, only the input terminal and output terminal
information of the power converters are used as
inputs to the ANN, while the capacitance is the ANN’s
target, and then the network is responsible for
estimating the value of C when using different inputs
than the trained ones. Normally, during the operation
of the power converter, the required terminal
information to train the ANN is supposed to be
available. Taking the power level of the applied
converter into consideration while the network is
trained, is improving the ANN’s estimated results by
being more robust against dynamic variations of the
loading power. Fig.3 illustrates the structure of the
proposed ANN. The basic structure of any neural
network consists of three layers, input, hidden, and
output layers. The input layer is where the available
amount of data N fed to the ANN will be stored. The
Fig .4.
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
hidden layer job is to transform the inputs into a
function that the output layer can use, while the
output layer transforms the hidden layer activations
into a scale which the operator wanted the output to
be on target.
III. CAPACITANCE ESTIMATION BASED ON
ANN
Capacitance estimation based on ANN is applied to a
back to back converter as shown in Fig. 4. The
specifications of the converter are listed in Table I.
Capacitance values in the range between (1000µF
and 5000µF) with 100µF step are used as targets to
the network, each value of these 41 samples
corresponds respectively to the single phase RMS
input/output voltages, currents, and the DC-link
voltage. Since the different loading condition of the
back-to-back converter is also considered, three
groups of 41 samples under the respective loading
level of 10 kW, 7 kW and 4 kW are used.
A back-to-back converter.
Table 1. THE SPECIFICATIONS OF THE BACK-TO-BACK CONVERTER
PARAMETERS.
Input AC Voltage (VL−L)
600 V
Output AC Voltage (VL−L)
380 V
Rated DC-link Voltage (Vdc)
780 V
Full Power Level (Po)
10 kW
Capacitance (C)
5000 µF
Fig .5.
Regression responce of the trained network..
296
Journal of Renewable Energy and Sustainable Development (RESD)
All the 123 samples are fed to a single hidden layer
ANN consisting of 10 neurons using the Neural Fitting
Tool (nftool) in MATLAB software. This tool is usually
used for estimation and prediction of problems in
which the neural network maps between a data set of
numeric inputs and a set of numeric targets. The
iteration algorithm used in this training is LevenbergMarquardt, which typically takes more memory but
less time. The training automatically stops when
generalisation stops improving, as indicated by an
increase in the mean square error of the validation
samples. During the training observing the
Regression value R is important, since Regression
values measure the correlation between outputs and
targets. An R value of 1 means a close
relationship, while 0 means a random relationship.
Fig. 5 shows the regression response of the trained
network in this paper. It can be noted that all the
123 input data are exactly aligned on the fitting line
and the values of R are close to 1. Based on this
result, the network stopped training and the ANN is
generated to be used.
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
Moreover, to prove the accuracy of the trained ANN,
the network is tested to identify the reduction of 50µF
out of 5000µF and the resulted estimation is shown in
Fig. 7. For the initial stage, the estimated value is
4991µF and for the degraded case, the estimated
value is 4945µF, which gives a 0.18% error as a
maximum.
To test the robustness of the trained ANN against
loading power variations, the ANN is tested to
estimate a capacitance
Fig .6.
Tested DC-link capacitor change in a back-to-back
converter.
IV. SIMULATION RESULTS AND DISCUSSION
In this section, the trained network is tested for
verification purpose. The inputs to the ANN are
stored in the Matlab work- space, and they are sent
to the ANN as a group set every 0.2 seconds. Each
group set is resulting in one corresponding output
(estimated C value). The same group set of inputs is
sent until a new set is available in the work-space.
The same back to back converter as shown in Fig. 4
is used for the simulation test, but with two capacitors
C1 and C2 connected in parallel through switches to
have the option to switch between them to simulate
the degradation of C. At the timing (5 sec) the
capacitor C2 will be switched on instead of C1 . Fig.6
shows the capacitance value estimated by the trained
ANN. The simulation results for the estimated
parameters and their corresponding errors are shown
in Table II. The estimated results verify that the
trained ANN is responding for the changes in the
capacitance value, and the statues of the capacitor
could be easily identified.
Table 2. SIMULATION RESULTS FOR ESTIMATED CAPACITANCE..
C1actual=5000µF
C1estimated=4991µ
F
Error=0.18%
C2actual=4000µF
C2estimated=3996µ
F
Error=0.1%
Fig .7.
Fig .8.
Trained ANN accuracy for the capacitance change shown
in Fig. 6.
Trained ANN accuracy for the loading power change.
value of 5000µF during a loading power change. The
esti- mated results with their corresponding errors
percentage are shown in Fig. 8.
297
Journal of Renewable Energy and Sustainable Development (RESD)
Moreover, for further verification of the ANN
accuracy, a set of random values of capacitance
between the ranges of (1000µF-5000µF) are applied
under three different power levels, their actual and
estimated values are shown in Fig.9. The actual and
estimated values of capacitance of the set applied
under 70% power level are presented in Table III.
In sake of showing the impact of the training data
amount on the accuracy of the trained ANN, another
network (ANN2) is trained by using 63 samples
instead of 123 samples considering the same
conditions of (ANN1) which have been trained earlier
in this paper. Fig. 10 shows that the errors estimated
by ANN2 are higher than the ones estimated
previously by ANN1, implying the trade-off between
estimation accuracy and required computation
resource.
Fig .9.
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
The last accuracy analysis is performed to observe
the error percentage of the estimated capacitance
with respect to different degree of changes of the
original value of 5000µF, the results are shown in Fig.
11. It can be noted that the estimation errors of the
proposed ANN are below 0.25%. It can respond and
estimate correctly the capacitance values even under
a very low level of capacitance reduction of 0.2%
changes.
The following remarks are given from the results
presented in this section:

The simulation results of the proposed method
based on ANN verify that condition monitoring
methods based on software solutions could be an
attractive alternative for the practical industry
applications.

It can be noted from the results that trained ANN
is capable to respond to a very small change of
capacitance
Estimation error analysis under different level of power.
Table 3. SIMULATION RESULTS FOR ESTIMATED CAPACITANCE (AT
7KW).
Fig .11. Estimation error analysis under different level of
capacitance reduction at rated power level.
Cactual=1300µF
Cestimated=1296µF
Error=0.3%
Cactual=1743µF
Cestimated=1747µF
Error=0.23%
Cactual=2600µF
Cestimated=2589µF
Error=0.4%
Cactual=3200µF
Cestimated=3203µF
Error=0.09%
and estimate the capacitance value within the range
in which the network is trained.
Cactual=3854µF
Cestimated=3850µF
Error=0.1%

Cactual=4265µF
Cestimated=4257µF
Error=0.18%
It should be noted that the accuracy of the trained
ANN strongly depends on the amount, quality,
and accuracy of the data used in the training.
V. CONCLUSIONS
Fig .10.
Estimation error analysis by different trained ANNs.
A new capacitor condition monitoring method based
on Artificial Neural Network algorithm is proposed in
this paper. It is applied to a back-to-back converter
study case to estimate the capacitance value change
of the DC-link capacitor. The proposed method
requires no additional hardware circuit and could be
implemented by using the spare resources of existing
298
Journal of Renewable Energy and Sustainable Development (RESD)
digital controllers in most of power electronic systems,
implying a minimum increased cost (e.g., only in the
research and development part). The error analysis
under different DC-link capacitance values and
different level of capacitance reduction with respect to
the initial value are given, achieving a maximum
estimation error that is well below 0.5%, which could
be acceptable in many practical applications. The
impact of training data amount on the error analysis is
also given.
REFERENCES
[1] S. Yang, D. Xiang, A. Bryant, P. Mawby, L. Ran,
and P. Tavner, “Condition monitoring for device
reliability in power electronic converters: A review,”
IEEE Transactions on Power Electronics, vol. 25, no.
11, pp. 2734–2752, Nov 2010.
Volume 1, Issue 2, December 2015 - ISSN 2356-8569
the condition monitoring of capacitors in power
electronic converters,” in Electromotion Joint International Conference (ACEMP - OPTIM), 2015 IEEE,
August 2015, pp. 243–249.
[5] D.-C. Lee, K.-J. Lee, J.-K. Seok, and J.-W. Choi,
“Online capacitance estimation of dc-link electrolytic
capacitors for three-phase ac/dc/ac pwm converters
using recursive least squares method,” in Proceedings
of Electric Power Applications, vol. 152, no. 6, pp.
1503–1508, Nov 2005.
[6] Y.-J. Jo, T. H. Nguyen, and D.-C. Lee, “Condition
monitoring of submodule capacitors in modular multilevel converters,” in Energy Conversion Congress
and Exposition (ECCE), 2014 IEEE, Sept 2014, pp.
2121–2126.
[7] T. H. Nguyen and D.-C. Lee, “Deterioration mon- itoring
of dc-link capacitors in ac machine drives by current
injection,” IEEE Transactions on Power Electronics,
vol. 30, no. 3, pp. 1126–1130, March
2015.
[2] X.-S. Pu, T. H. Nguyen, D.-C. Lee, K.-B. Lee, and J.-M.
Kim, “Fault diagnosis of dc-link capacitors in threephase ac/dc pwm converters by online estima- tion of
equivalent series resistance,” IEEE Transac- tions on
Industrial Electronics, vol. 60, no. 9, pp. 4118–4127,
Sept 2013.
[8] A. Wechsler, B. Mecrow, D. Atkinson, J. Bennett, and
M. Benarous, “Condition monitoring of dc-link
capacitors in aerospace drives,” IEEE Transactions on
Industry Applications, vol. 48, no. 6, pp. 1866– 1874,
Nov 2012.
[3] H. Wang and F. Blaabjerg, “Reliability of capac- itors
for dc-link applications in power electronic
converters-an overview,” IEEE Transactions on Industry Applications, vol. 50, no. 5, pp. 3569–3578,
Sept 2014.
[9] A. Abo-Khalil and D.-C. Lee, “Dc-link capacitance
estimation in ac/dc/ac pwm converters using voltage
injection,” IEEE Transactions on Industry Applications, vol. 44, no. 5, pp. 1631–1637, Sept 2008.
[4] H. Soliman, H. Wang, and F. Blaabjerg, “A review of
299